kaorumuta / isic2016_challenge Goto Github PK

Hello, KaoruMuta. Thx for your very nice project on the ISIC 2016 dataset. I have been working on using my own CNN architecture to test on ISIC 2016 dataset. Therefore, I can not use the pre-trained weights from the ImageNet dataset. I did experiments by comparing using or not using pre-trained weight of ResNet50 on ISIC 2016 dataset, while the difference is very large (not using pre-trained weight is much worse). Have you ever tried this before? I think I should first use non-pre-trained ResNet50 with the training parameter settings to get good results on this dataset, and then I can use my own designed architecture. I meet this problem now, could you please give me some suggestions.

Your open source content is very good. but I met some questions when I try to run the program according to your instructions，I find that my results are generally different from any experimental results in your result directory. I run the train_ISIC2016.py program with trainModel、 batchSize = 32、trainingEpochs=20、 plt、 cutout and focal settings.Some of the results are as follows：




accuracy.txt
I don't know what the problem is, so can you give me some advice? Thank you very much！